X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/f04cfcce18c6e4bd1faf3140ccbc835328580024..293631794d64c64986ba0db2568345c005c90790:/services/keepstore/s3_volume.go diff --git a/services/keepstore/s3_volume.go b/services/keepstore/s3_volume.go index 79a680d58a..2e2e97a974 100644 --- a/services/keepstore/s3_volume.go +++ b/services/keepstore/s3_volume.go @@ -1,330 +1,845 @@ -package main +// Copyright (C) The Arvados Authors. All rights reserved. +// +// SPDX-License-Identifier: AGPL-3.0 + +package keepstore import ( + "bytes" + "context" "encoding/base64" "encoding/hex" - "flag" + "encoding/json" + "errors" "fmt" "io" - "log" - "net/http" "os" "regexp" + "strings" + "sync" + "sync/atomic" "time" - "github.com/AdRoll/goamz/aws" - "github.com/AdRoll/goamz/s3" + "git.arvados.org/arvados.git/sdk/go/arvados" + "github.com/aws/aws-sdk-go-v2/aws" + "github.com/aws/aws-sdk-go-v2/aws/awserr" + "github.com/aws/aws-sdk-go-v2/aws/defaults" + "github.com/aws/aws-sdk-go-v2/aws/ec2metadata" + "github.com/aws/aws-sdk-go-v2/aws/ec2rolecreds" + "github.com/aws/aws-sdk-go-v2/aws/endpoints" + "github.com/aws/aws-sdk-go-v2/service/s3" + "github.com/aws/aws-sdk-go-v2/service/s3/s3manager" + "github.com/prometheus/client_golang/prometheus" + "github.com/sirupsen/logrus" ) -var ( - ErrS3DeleteNotAvailable = fmt.Errorf("delete without -s3-unsafe-delete is not implemented") - - s3AccessKeyFile string - s3SecretKeyFile string - s3RegionName string - s3Endpoint string - s3Replication int - s3UnsafeDelete bool +func init() { + driver["S3"] = news3Volume +} - s3ACL = s3.Private +const ( + s3DefaultReadTimeout = arvados.Duration(10 * time.Minute) + s3DefaultConnectTimeout = arvados.Duration(time.Minute) + maxClockSkew = 600 * time.Second + nearlyRFC1123 = "Mon, 2 Jan 2006 15:04:05 GMT" + s3downloaderPartSize = 6 * 1024 * 1024 + s3downloaderReadConcurrency = 11 + s3uploaderPartSize = 5 * 1024 * 1024 + s3uploaderWriteConcurrency = 5 ) -const ( - maxClockSkew = 600 * time.Second - nearlyRFC1123 = "Mon, 2 Jan 2006 15:04:05 GMT" +var ( + errS3TrashDisabled = fmt.Errorf("trash function is disabled because Collections.BlobTrashLifetime=0 and DriverParameters.UnsafeDelete=false") + s3AWSKeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`) + s3AWSZeroTime time.Time ) -type s3VolumeAdder struct { - *volumeSet +// s3Volume implements Volume using an S3 bucket. +type s3Volume struct { + arvados.S3VolumeDriverParameters + AuthToken string // populated automatically when IAMRole is used + AuthExpiration time.Time // populated automatically when IAMRole is used + + cluster *arvados.Cluster + volume arvados.Volume + logger logrus.FieldLogger + metrics *volumeMetricsVecs + bufferPool *bufferPool + bucket *s3Bucket + region string + startOnce sync.Once } -func (s *s3VolumeAdder) Set(bucketName string) error { - if trashLifetime != 0 { - return ErrNotImplemented - } - if bucketName == "" { - return fmt.Errorf("no container name given") - } - if s3AccessKeyFile == "" || s3SecretKeyFile == "" { - return fmt.Errorf("-s3-access-key-file and -s3-secret-key-file arguments must given before -s3-bucket-volume") +// s3bucket wraps s3.bucket and counts I/O and API usage stats. The +// wrapped bucket can be replaced atomically with SetBucket in order +// to update credentials. +type s3Bucket struct { + bucket string + svc *s3.Client + stats s3awsbucketStats + mu sync.Mutex +} + +func (v *s3Volume) isKeepBlock(s string) (string, bool) { + if v.PrefixLength > 0 && len(s) == v.PrefixLength+33 && s[:v.PrefixLength] == s[v.PrefixLength+1:v.PrefixLength*2+1] { + s = s[v.PrefixLength+1:] } - region, ok := aws.Regions[s3RegionName] - if s3Endpoint == "" { - if !ok { - return fmt.Errorf("unrecognized region %+q; try specifying -s3-endpoint instead", s3RegionName) - } + return s, s3AWSKeepBlockRegexp.MatchString(s) +} + +// Return the key used for a given loc. If PrefixLength==0 then +// key("abcdef0123") is "abcdef0123", if PrefixLength==3 then key is +// "abc/abcdef0123", etc. +func (v *s3Volume) key(loc string) string { + if v.PrefixLength > 0 && v.PrefixLength < len(loc)-1 { + return loc[:v.PrefixLength] + "/" + loc } else { - if ok { - return fmt.Errorf("refusing to use AWS region name %+q with endpoint %+q; "+ - "specify empty endpoint (\"-s3-endpoint=\") or use a different region name", s3RegionName, s3Endpoint) - } - region = aws.Region{ - Name: s3RegionName, - S3Endpoint: s3Endpoint, - } + return loc } - var err error - var auth aws.Auth - auth.AccessKey, err = readKeyFromFile(s3AccessKeyFile) - if err != nil { - return err +} + +func news3Volume(params newVolumeParams) (volume, error) { + v := &s3Volume{ + cluster: params.Cluster, + volume: params.ConfigVolume, + metrics: params.MetricsVecs, + bufferPool: params.BufferPool, } - auth.SecretKey, err = readKeyFromFile(s3SecretKeyFile) + err := json.Unmarshal(params.ConfigVolume.DriverParameters, v) if err != nil { - return err + return nil, err + } + v.logger = params.Logger.WithField("Volume", v.DeviceID()) + return v, v.check("") +} + +func (v *s3Volume) translateError(err error) error { + if _, ok := err.(*aws.RequestCanceledError); ok { + return context.Canceled + } else if aerr, ok := err.(awserr.Error); ok { + if aerr.Code() == "NotFound" { + return os.ErrNotExist + } else if aerr.Code() == "NoSuchKey" { + return os.ErrNotExist + } } - if flagSerializeIO { - log.Print("Notice: -serialize is not supported by s3-bucket volumes.") + return err +} + +// safeCopy calls CopyObjectRequest, and checks the response to make +// sure the copy succeeded and updated the timestamp on the +// destination object +// +// (If something goes wrong during the copy, the error will be +// embedded in the 200 OK response) +func (v *s3Volume) safeCopy(dst, src string) error { + input := &s3.CopyObjectInput{ + Bucket: aws.String(v.bucket.bucket), + ContentType: aws.String("application/octet-stream"), + CopySource: aws.String(v.bucket.bucket + "/" + src), + Key: aws.String(dst), } - v := NewS3Volume(auth, region, bucketName, flagReadonly, s3Replication) - if err := v.Check(); err != nil { + + req := v.bucket.svc.CopyObjectRequest(input) + resp, err := req.Send(context.Background()) + + err = v.translateError(err) + if os.IsNotExist(err) { return err + } else if err != nil { + return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.bucket+"/"+src, err) + } + + if resp.CopyObjectResult.LastModified == nil { + return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.CopyObjectResult.LastModified, err) + } else if time.Now().Sub(*resp.CopyObjectResult.LastModified) > maxClockSkew { + return fmt.Errorf("PutCopy succeeded but returned an old timestamp: %q: %s", resp.CopyObjectResult.LastModified, resp.CopyObjectResult.LastModified) } - *s.volumeSet = append(*s.volumeSet, v) return nil } -func s3regions() (okList []string) { - for r, _ := range aws.Regions { - okList = append(okList, r) +func (v *s3Volume) check(ec2metadataHostname string) error { + if v.Bucket == "" { + return errors.New("DriverParameters: Bucket must be provided") + } + if v.IndexPageSize == 0 { + v.IndexPageSize = 1000 + } + if v.RaceWindow < 0 { + return errors.New("DriverParameters: RaceWindow must not be negative") } - return -} -func init() { - flag.Var(&s3VolumeAdder{&volumes}, - "s3-bucket-volume", - "Use the given bucket as a storage volume. Can be given multiple times.") - flag.StringVar( - &s3RegionName, - "s3-region", - "", - fmt.Sprintf("AWS region used for subsequent -s3-bucket-volume arguments. Allowed values are %+q.", s3regions())) - flag.StringVar( - &s3Endpoint, - "s3-endpoint", - "", - "Endpoint URL used for subsequent -s3-bucket-volume arguments. If blank, use the AWS endpoint corresponding to the -s3-region argument. For Google Storage, use \"https://storage.googleapis.com\".") - flag.StringVar( - &s3AccessKeyFile, - "s3-access-key-file", - "", - "File containing the access key used for subsequent -s3-bucket-volume arguments.") - flag.StringVar( - &s3SecretKeyFile, - "s3-secret-key-file", - "", - "File containing the secret key used for subsequent -s3-bucket-volume arguments.") - flag.IntVar( - &s3Replication, - "s3-replication", - 2, - "Replication level reported to clients for subsequent -s3-bucket-volume arguments.") - flag.BoolVar( - &s3UnsafeDelete, - "s3-unsafe-delete", - false, - "EXPERIMENTAL. Enable deletion (garbage collection), even though there are known race conditions that can cause data loss.") -} - -type S3Volume struct { - *s3.Bucket - readonly bool - replication int - indexPageSize int -} - -// NewS3Volume returns a new S3Volume using the given auth, region, -// and bucket name. The replication argument specifies the replication -// level to report when writing data. -func NewS3Volume(auth aws.Auth, region aws.Region, bucket string, readonly bool, replication int) *S3Volume { - return &S3Volume{ - Bucket: &s3.Bucket{ - S3: s3.New(auth, region), - Name: bucket, - }, - readonly: readonly, - replication: replication, - indexPageSize: 1000, - } -} - -func (v *S3Volume) Check() error { + if v.V2Signature { + return errors.New("DriverParameters: V2Signature is not supported") + } + + defaultResolver := endpoints.NewDefaultResolver() + + cfg := defaults.Config() + + if v.Endpoint == "" && v.Region == "" { + return fmt.Errorf("AWS region or endpoint must be specified") + } else if v.Endpoint != "" || ec2metadataHostname != "" { + myCustomResolver := func(service, region string) (aws.Endpoint, error) { + if v.Endpoint != "" && service == "s3" { + return aws.Endpoint{ + URL: v.Endpoint, + SigningRegion: region, + }, nil + } else if service == "ec2metadata" && ec2metadataHostname != "" { + return aws.Endpoint{ + URL: ec2metadataHostname, + }, nil + } else { + return defaultResolver.ResolveEndpoint(service, region) + } + } + cfg.EndpointResolver = aws.EndpointResolverFunc(myCustomResolver) + } + if v.Region == "" { + // Endpoint is already specified (otherwise we would + // have errored out above), but Region is also + // required by the aws sdk, in order to determine + // SignatureVersions. + v.Region = "us-east-1" + } + cfg.Region = v.Region + + // Zero timeouts mean "wait forever", which is a bad + // default. Default to long timeouts instead. + if v.ConnectTimeout == 0 { + v.ConnectTimeout = s3DefaultConnectTimeout + } + if v.ReadTimeout == 0 { + v.ReadTimeout = s3DefaultReadTimeout + } + + creds := aws.NewChainProvider( + []aws.CredentialsProvider{ + aws.NewStaticCredentialsProvider(v.AccessKeyID, v.SecretAccessKey, v.AuthToken), + ec2rolecreds.New(ec2metadata.New(cfg), func(opts *ec2rolecreds.ProviderOptions) { + // (from aws-sdk-go-v2 comments) + // "allow the credentials to trigger + // refreshing prior to the credentials + // actually expiring. This is + // beneficial so race conditions with + // expiring credentials do not cause + // request to fail unexpectedly due to + // ExpiredTokenException exceptions." + // + // (from + // https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/iam-roles-for-amazon-ec2.html) + // "We make new credentials available + // at least five minutes before the + // expiration of the old credentials." + opts.ExpiryWindow = 5 * time.Minute + }), + }) + + cfg.Credentials = creds + + v.bucket = &s3Bucket{ + bucket: v.Bucket, + svc: s3.New(cfg), + } + + // Set up prometheus metrics + lbls := prometheus.Labels{"device_id": v.DeviceID()} + v.bucket.stats.opsCounters, v.bucket.stats.errCounters, v.bucket.stats.ioBytes = v.metrics.getCounterVecsFor(lbls) + return nil } -func (v *S3Volume) Get(loc string) ([]byte, error) { - rdr, err := v.Bucket.GetReader(loc) - if err != nil { - return nil, v.translateError(err) +// DeviceID returns a globally unique ID for the storage bucket. +func (v *s3Volume) DeviceID() string { + return "s3://" + v.Endpoint + "/" + v.Bucket +} + +// EmptyTrash looks for trashed blocks that exceeded BlobTrashLifetime +// and deletes them from the volume. +func (v *s3Volume) EmptyTrash() { + var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64 + + // Define "ready to delete" as "...when EmptyTrash started". + startT := time.Now() + + emptyOneKey := func(trash *s3.Object) { + key := strings.TrimPrefix(*trash.Key, "trash/") + loc, isblk := v.isKeepBlock(key) + if !isblk { + return + } + atomic.AddInt64(&bytesInTrash, *trash.Size) + atomic.AddInt64(&blocksInTrash, 1) + + trashT := *trash.LastModified + recent, err := v.head("recent/" + key) + if err != nil && os.IsNotExist(v.translateError(err)) { + v.logger.Warnf("EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", *trash.Key, "recent/"+key, err) + err = v.BlockUntrash(loc) + if err != nil { + v.logger.WithError(err).Errorf("EmptyTrash: Untrash(%q) failed", loc) + } + return + } else if err != nil { + v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", "recent/"+key) + return + } + if trashT.Sub(*recent.LastModified) < v.cluster.Collections.BlobSigningTTL.Duration() { + if age := startT.Sub(*recent.LastModified); age >= v.cluster.Collections.BlobSigningTTL.Duration()-time.Duration(v.RaceWindow) { + // recent/key is too old to protect + // loc from being Trashed again during + // the raceWindow that starts if we + // delete trash/X now. + // + // Note this means (TrashSweepInterval + // < BlobSigningTTL - raceWindow) is + // necessary to avoid starvation. + v.logger.Infof("EmptyTrash: detected old race for %q, calling fixRace + Touch", loc) + v.fixRace(key) + v.BlockTouch(loc) + return + } + _, err := v.head(key) + if os.IsNotExist(err) { + v.logger.Infof("EmptyTrash: detected recent race for %q, calling fixRace", loc) + v.fixRace(key) + return + } else if err != nil { + v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", loc) + return + } + } + if startT.Sub(trashT) < v.cluster.Collections.BlobTrashLifetime.Duration() { + return + } + err = v.bucket.Del(*trash.Key) + if err != nil { + v.logger.WithError(err).Errorf("EmptyTrash: error deleting %q", *trash.Key) + return + } + atomic.AddInt64(&bytesDeleted, *trash.Size) + atomic.AddInt64(&blocksDeleted, 1) + + _, err = v.head(*trash.Key) + if err == nil { + v.logger.Warnf("EmptyTrash: HEAD %q succeeded immediately after deleting %q", loc, loc) + return + } + if !os.IsNotExist(v.translateError(err)) { + v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", key) + return + } + err = v.bucket.Del("recent/" + key) + if err != nil { + v.logger.WithError(err).Warnf("EmptyTrash: error deleting %q", "recent/"+key) + } } - defer rdr.Close() - buf := bufs.Get(BlockSize) - n, err := io.ReadFull(rdr, buf) - switch err { - case nil, io.EOF, io.ErrUnexpectedEOF: - return buf[:n], nil - default: - bufs.Put(buf) - return nil, v.translateError(err) + + var wg sync.WaitGroup + todo := make(chan *s3.Object, v.cluster.Collections.BlobDeleteConcurrency) + for i := 0; i < v.cluster.Collections.BlobDeleteConcurrency; i++ { + wg.Add(1) + go func() { + defer wg.Done() + for key := range todo { + emptyOneKey(key) + } + }() + } + + trashL := s3awsLister{ + Logger: v.logger, + Bucket: v.bucket, + Prefix: "trash/", + PageSize: v.IndexPageSize, + Stats: &v.bucket.stats, + } + for trash := trashL.First(); trash != nil; trash = trashL.Next() { + todo <- trash + } + close(todo) + wg.Wait() + + if err := trashL.Error(); err != nil { + v.logger.WithError(err).Error("EmptyTrash: lister failed") } + v.logger.Infof("EmptyTrash: stats for %v: Deleted %v bytes in %v blocks. Remaining in trash: %v bytes in %v blocks.", v.DeviceID(), bytesDeleted, blocksDeleted, bytesInTrash-bytesDeleted, blocksInTrash-blocksDeleted) } -func (v *S3Volume) Compare(loc string, expect []byte) error { - rdr, err := v.Bucket.GetReader(loc) +// fixRace(X) is called when "recent/X" exists but "X" doesn't +// exist. If the timestamps on "recent/X" and "trash/X" indicate there +// was a race between Put and Trash, fixRace recovers from the race by +// Untrashing the block. +func (v *s3Volume) fixRace(key string) bool { + trash, err := v.head("trash/" + key) if err != nil { - return v.translateError(err) + if !os.IsNotExist(v.translateError(err)) { + v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "trash/"+key) + } + return false + } + + recent, err := v.head("recent/" + key) + if err != nil { + v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "recent/"+key) + return false + } + + recentTime := *recent.LastModified + trashTime := *trash.LastModified + ageWhenTrashed := trashTime.Sub(recentTime) + if ageWhenTrashed >= v.cluster.Collections.BlobSigningTTL.Duration() { + // No evidence of a race: block hasn't been written + // since it became eligible for Trash. No fix needed. + return false } - defer rdr.Close() - return v.translateError(compareReaderWithBuf(rdr, expect, loc[:32])) + + v.logger.Infof("fixRace: %q: trashed at %s but touched at %s (age when trashed = %s < %s)", key, trashTime, recentTime, ageWhenTrashed, v.cluster.Collections.BlobSigningTTL) + v.logger.Infof("fixRace: copying %q to %q to recover from race between Put/Touch and Trash", "recent/"+key, key) + err = v.safeCopy(key, "trash/"+key) + if err != nil { + v.logger.WithError(err).Error("fixRace: copy failed") + return false + } + return true } -func (v *S3Volume) Put(loc string, block []byte) error { - if v.readonly { - return MethodDisabledError +func (v *s3Volume) head(key string) (result *s3.HeadObjectOutput, err error) { + input := &s3.HeadObjectInput{ + Bucket: aws.String(v.bucket.bucket), + Key: aws.String(key), } - var opts s3.Options - if len(block) > 0 { - md5, err := hex.DecodeString(loc) + + req := v.bucket.svc.HeadObjectRequest(input) + res, err := req.Send(context.TODO()) + + v.bucket.stats.TickOps("head") + v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.HeadOps) + v.bucket.stats.TickErr(err) + + if err != nil { + return nil, v.translateError(err) + } + result = res.HeadObjectOutput + return +} + +// BlockRead reads a Keep block that has been stored as a block blob +// in the S3 bucket. +func (v *s3Volume) BlockRead(ctx context.Context, hash string, w io.WriterAt) error { + key := v.key(hash) + err := v.readWorker(ctx, key, w) + if err != nil { + err = v.translateError(err) + if !os.IsNotExist(err) { + return err + } + + _, err = v.head("recent/" + key) + err = v.translateError(err) + if err != nil { + // If we can't read recent/X, there's no point in + // trying fixRace. Give up. + return err + } + if !v.fixRace(key) { + err = os.ErrNotExist + return err + } + + err = v.readWorker(ctx, key, w) if err != nil { + v.logger.Warnf("reading %s after successful fixRace: %s", hash, err) + err = v.translateError(err) return err } - opts.ContentMD5 = base64.StdEncoding.EncodeToString(md5) } - return v.translateError( - v.Bucket.Put( - loc, block, "application/octet-stream", s3ACL, opts)) + return nil } -func (v *S3Volume) Touch(loc string) error { - if v.readonly { - return MethodDisabledError +func (v *s3Volume) readWorker(ctx context.Context, key string, dst io.WriterAt) error { + downloader := s3manager.NewDownloaderWithClient(v.bucket.svc, func(u *s3manager.Downloader) { + u.PartSize = s3downloaderPartSize + u.Concurrency = s3downloaderReadConcurrency + }) + count, err := downloader.DownloadWithContext(ctx, dst, &s3.GetObjectInput{ + Bucket: aws.String(v.bucket.bucket), + Key: aws.String(key), + }) + v.bucket.stats.TickOps("get") + v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.GetOps) + v.bucket.stats.TickErr(err) + v.bucket.stats.TickInBytes(uint64(count)) + return v.translateError(err) +} + +func (v *s3Volume) writeObject(ctx context.Context, key string, r io.Reader) error { + if r == nil { + // r == nil leads to a memory violation in func readFillBuf in + // aws-sdk-go-v2@v0.23.0/service/s3/s3manager/upload.go + r = bytes.NewReader(nil) } - result, err := v.Bucket.PutCopy(loc, s3ACL, s3.CopyOptions{ - ContentType: "application/octet-stream", - MetadataDirective: "REPLACE", - }, v.Bucket.Name+"/"+loc) - if err != nil { - return v.translateError(err) + + uploadInput := s3manager.UploadInput{ + Bucket: aws.String(v.bucket.bucket), + Key: aws.String(key), + Body: r, } - t, err := time.Parse(time.RFC3339, result.LastModified) + + if loc, ok := v.isKeepBlock(key); ok { + var contentMD5 string + md5, err := hex.DecodeString(loc) + if err != nil { + return v.translateError(err) + } + contentMD5 = base64.StdEncoding.EncodeToString(md5) + uploadInput.ContentMD5 = &contentMD5 + } + + // Experimentation indicated that using concurrency 5 yields the best + // throughput, better than higher concurrency (10 or 13) by ~5%. + // Defining u.BufferProvider = s3manager.NewBufferedReadSeekerWriteToPool(64 * 1024 * 1024) + // is detrimental to throughput (minus ~15%). + uploader := s3manager.NewUploaderWithClient(v.bucket.svc, func(u *s3manager.Uploader) { + u.PartSize = s3uploaderPartSize + u.Concurrency = s3uploaderWriteConcurrency + }) + + // Unlike the goamz S3 driver, we don't need to precompute ContentSHA256: + // the aws-sdk-go v2 SDK uses a ReadSeeker to avoid having to copy the + // block, so there is no extra memory use to be concerned about. See + // makeSha256Reader in aws/signer/v4/v4.go. In fact, we explicitly disable + // calculating the Sha-256 because we don't need it; we already use md5sum + // hashes that match the name of the block. + _, err := uploader.UploadWithContext(ctx, &uploadInput, s3manager.WithUploaderRequestOptions(func(r *aws.Request) { + r.HTTPRequest.Header.Set("X-Amz-Content-Sha256", "UNSIGNED-PAYLOAD") + })) + + v.bucket.stats.TickOps("put") + v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.PutOps) + v.bucket.stats.TickErr(err) + + return v.translateError(err) +} + +// Put writes a block. +func (v *s3Volume) BlockWrite(ctx context.Context, hash string, data []byte) error { + // Do not use putWithPipe here; we want to pass an io.ReadSeeker to the S3 + // sdk to avoid memory allocation there. See #17339 for more information. + rdr := bytes.NewReader(data) + r := newCountingReaderAtSeeker(rdr, v.bucket.stats.TickOutBytes) + key := v.key(hash) + err := v.writeObject(ctx, key, r) if err != nil { return err } - if time.Since(t) > maxClockSkew { - return fmt.Errorf("PutCopy returned old LastModified %s => %s (%s ago)", result.LastModified, t, time.Since(t)) + return v.writeObject(ctx, "recent/"+key, nil) +} + +type s3awsLister struct { + Logger logrus.FieldLogger + Bucket *s3Bucket + Prefix string + PageSize int + Stats *s3awsbucketStats + ContinuationToken string + buf []s3.Object + err error +} + +// First fetches the first page and returns the first item. It returns +// nil if the response is the empty set or an error occurs. +func (lister *s3awsLister) First() *s3.Object { + lister.getPage() + return lister.pop() +} + +// Next returns the next item, fetching the next page if necessary. It +// returns nil if the last available item has already been fetched, or +// an error occurs. +func (lister *s3awsLister) Next() *s3.Object { + if len(lister.buf) == 0 && lister.ContinuationToken != "" { + lister.getPage() } - return nil + return lister.pop() +} + +// Return the most recent error encountered by First or Next. +func (lister *s3awsLister) Error() error { + return lister.err } -func (v *S3Volume) Mtime(loc string) (time.Time, error) { - resp, err := v.Bucket.Head(loc, nil) +func (lister *s3awsLister) getPage() { + lister.Stats.TickOps("list") + lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps) + + var input *s3.ListObjectsV2Input + if lister.ContinuationToken == "" { + input = &s3.ListObjectsV2Input{ + Bucket: aws.String(lister.Bucket.bucket), + MaxKeys: aws.Int64(int64(lister.PageSize)), + Prefix: aws.String(lister.Prefix), + } + } else { + input = &s3.ListObjectsV2Input{ + Bucket: aws.String(lister.Bucket.bucket), + MaxKeys: aws.Int64(int64(lister.PageSize)), + Prefix: aws.String(lister.Prefix), + ContinuationToken: &lister.ContinuationToken, + } + } + + req := lister.Bucket.svc.ListObjectsV2Request(input) + resp, err := req.Send(context.Background()) if err != nil { - return zeroTime, v.translateError(err) + if aerr, ok := err.(awserr.Error); ok { + lister.err = aerr + } else { + lister.err = err + } + return } - hdr := resp.Header.Get("Last-Modified") - t, err := time.Parse(time.RFC1123, hdr) - if err != nil && hdr != "" { - // AWS example is "Sun, 1 Jan 2006 12:00:00 GMT", - // which isn't quite "Sun, 01 Jan 2006 12:00:00 GMT" - // as required by HTTP spec. If it's not a valid HTTP - // header value, it's probably AWS (or s3test) giving - // us a nearly-RFC1123 timestamp. - t, err = time.Parse(nearlyRFC1123, hdr) + + if *resp.IsTruncated { + lister.ContinuationToken = *resp.NextContinuationToken + } else { + lister.ContinuationToken = "" + } + lister.buf = make([]s3.Object, 0, len(resp.Contents)) + for _, key := range resp.Contents { + if !strings.HasPrefix(*key.Key, lister.Prefix) { + lister.Logger.Warnf("s3awsLister: S3 Bucket.List(prefix=%q) returned key %q", lister.Prefix, *key.Key) + continue + } + lister.buf = append(lister.buf, key) } - return t, err } -func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error { - nextMarker := "" - for { - listResp, err := v.Bucket.List(prefix, "", nextMarker, v.indexPageSize) - if err != nil { - return err +func (lister *s3awsLister) pop() (k *s3.Object) { + if len(lister.buf) > 0 { + k = &lister.buf[0] + lister.buf = lister.buf[1:] + } + return +} + +// Index writes a complete list of locators with the given prefix +// for which Get() can retrieve data. +func (v *s3Volume) Index(ctx context.Context, prefix string, writer io.Writer) error { + prefix = v.key(prefix) + // Use a merge sort to find matching sets of X and recent/X. + dataL := s3awsLister{ + Logger: v.logger, + Bucket: v.bucket, + Prefix: prefix, + PageSize: v.IndexPageSize, + Stats: &v.bucket.stats, + } + recentL := s3awsLister{ + Logger: v.logger, + Bucket: v.bucket, + Prefix: "recent/" + prefix, + PageSize: v.IndexPageSize, + Stats: &v.bucket.stats, + } + for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() { + if ctx.Err() != nil { + return ctx.Err() } - for _, key := range listResp.Contents { - t, err := time.Parse(time.RFC3339, key.LastModified) - if err != nil { - return err - } - if !v.isKeepBlock(key.Key) { + if *data.Key >= "g" { + // Conveniently, "recent/*" and "trash/*" are + // lexically greater than all hex-encoded data + // hashes, so stopping here avoids iterating + // over all of them needlessly with dataL. + break + } + loc, isblk := v.isKeepBlock(*data.Key) + if !isblk { + continue + } + + // stamp is the list entry we should use to report the + // last-modified time for this data block: it will be + // the recent/X entry if one exists, otherwise the + // entry for the data block itself. + stamp := data + + // Advance to the corresponding recent/X marker, if any + for recent != nil && recentL.Error() == nil { + if cmp := strings.Compare((*recent.Key)[7:], *data.Key); cmp < 0 { + recent = recentL.Next() continue + } else if cmp == 0 { + stamp = recent + recent = recentL.Next() + break + } else { + // recent/X marker is missing: we'll + // use the timestamp on the data + // object. + break } - fmt.Fprintf(writer, "%s+%d %d\n", key.Key, key.Size, t.Unix()) } - if !listResp.IsTruncated { - break + if err := recentL.Error(); err != nil { + return err } - nextMarker = listResp.NextMarker + // We truncate sub-second precision here. Otherwise + // timestamps will never match the RFC1123-formatted + // Last-Modified values parsed by Mtime(). + fmt.Fprintf(writer, "%s+%d %d\n", loc, *data.Size, stamp.LastModified.Unix()*1000000000) } - return nil + return dataL.Error() } -// Trash a Keep block. -func (v *S3Volume) Trash(loc string) error { - if v.readonly { - return MethodDisabledError - } - if trashLifetime != 0 { - return ErrNotImplemented - } - if t, err := v.Mtime(loc); err != nil { - return err - } else if time.Since(t) < blobSignatureTTL { - return nil +// Mtime returns the stored timestamp for the given locator. +func (v *s3Volume) Mtime(loc string) (time.Time, error) { + key := v.key(loc) + _, err := v.head(key) + if err != nil { + return s3AWSZeroTime, v.translateError(err) } - if !s3UnsafeDelete { - return ErrS3DeleteNotAvailable + resp, err := v.head("recent/" + key) + err = v.translateError(err) + if os.IsNotExist(err) { + // The data object X exists, but recent/X is missing. + err = v.writeObject(context.Background(), "recent/"+key, nil) + if err != nil { + v.logger.WithError(err).Errorf("error creating %q", "recent/"+key) + return s3AWSZeroTime, v.translateError(err) + } + v.logger.Infof("Mtime: created %q to migrate existing block to new storage scheme", "recent/"+key) + resp, err = v.head("recent/" + key) + if err != nil { + v.logger.WithError(err).Errorf("HEAD failed after creating %q", "recent/"+key) + return s3AWSZeroTime, v.translateError(err) + } + } else if err != nil { + // HEAD recent/X failed for some other reason. + return s3AWSZeroTime, err } - return v.Bucket.Del(loc) + return *resp.LastModified, err } -// TBD -func (v *S3Volume) Untrash(loc string) error { - return ErrNotImplemented +// InternalStats returns bucket I/O and API call counters. +func (v *s3Volume) InternalStats() interface{} { + return &v.bucket.stats } -func (v *S3Volume) Status() *VolumeStatus { - return &VolumeStatus{ - DeviceNum: 1, - BytesFree: BlockSize * 1000, - BytesUsed: 1, +// BlockTouch sets the timestamp for the given locator to the current time. +func (v *s3Volume) BlockTouch(hash string) error { + key := v.key(hash) + _, err := v.head(key) + err = v.translateError(err) + if os.IsNotExist(err) && v.fixRace(key) { + // The data object got trashed in a race, but fixRace + // rescued it. + } else if err != nil { + return err } + err = v.writeObject(context.Background(), "recent/"+key, nil) + return v.translateError(err) } -func (v *S3Volume) String() string { - return fmt.Sprintf("s3-bucket:%+q", v.Bucket.Name) +// checkRaceWindow returns a non-nil error if trash/key is, or might +// be, in the race window (i.e., it's not safe to trash key). +func (v *s3Volume) checkRaceWindow(key string) error { + resp, err := v.head("trash/" + key) + err = v.translateError(err) + if os.IsNotExist(err) { + // OK, trash/X doesn't exist so we're not in the race + // window + return nil + } else if err != nil { + // Error looking up trash/X. We don't know whether + // we're in the race window + return err + } + t := resp.LastModified + safeWindow := t.Add(v.cluster.Collections.BlobTrashLifetime.Duration()).Sub(time.Now().Add(time.Duration(v.RaceWindow))) + if safeWindow <= 0 { + // We can't count on "touch trash/X" to prolong + // trash/X's lifetime. The new timestamp might not + // become visible until now+raceWindow, and EmptyTrash + // is allowed to delete trash/X before then. + return fmt.Errorf("%s: same block is already in trash, and safe window ended %s ago", key, -safeWindow) + } + // trash/X exists, but it won't be eligible for deletion until + // after now+raceWindow, so it's safe to overwrite it. + return nil } -func (v *S3Volume) Writable() bool { - return !v.readonly +func (b *s3Bucket) Del(path string) error { + input := &s3.DeleteObjectInput{ + Bucket: aws.String(b.bucket), + Key: aws.String(path), + } + req := b.svc.DeleteObjectRequest(input) + _, err := req.Send(context.Background()) + b.stats.TickOps("delete") + b.stats.Tick(&b.stats.Ops, &b.stats.DelOps) + b.stats.TickErr(err) + return err } -func (v *S3Volume) Replication() int { - return v.replication + +// Trash a Keep block. +func (v *s3Volume) BlockTrash(loc string) error { + if t, err := v.Mtime(loc); err != nil { + return err + } else if time.Since(t) < v.cluster.Collections.BlobSigningTTL.Duration() { + return nil + } + key := v.key(loc) + if v.cluster.Collections.BlobTrashLifetime == 0 { + if !v.UnsafeDelete { + return errS3TrashDisabled + } + return v.translateError(v.bucket.Del(key)) + } + err := v.checkRaceWindow(key) + if err != nil { + return err + } + err = v.safeCopy("trash/"+key, key) + if err != nil { + return err + } + return v.translateError(v.bucket.Del(key)) } -var s3KeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`) +// BlockUntrash moves block from trash back into store +func (v *s3Volume) BlockUntrash(hash string) error { + key := v.key(hash) + err := v.safeCopy(key, "trash/"+key) + if err != nil { + return err + } + err = v.writeObject(context.Background(), "recent/"+key, nil) + return v.translateError(err) +} -func (v *S3Volume) isKeepBlock(s string) bool { - return s3KeepBlockRegexp.MatchString(s) +type s3awsbucketStats struct { + statsTicker + Ops uint64 + GetOps uint64 + PutOps uint64 + HeadOps uint64 + DelOps uint64 + ListOps uint64 } -func (v *S3Volume) translateError(err error) error { - switch err := err.(type) { - case *s3.Error: - if err.StatusCode == http.StatusNotFound && err.Code == "NoSuchKey" { - return os.ErrNotExist +func (s *s3awsbucketStats) TickErr(err error) { + if err == nil { + return + } + errType := fmt.Sprintf("%T", err) + if aerr, ok := err.(awserr.Error); ok { + if reqErr, ok := err.(awserr.RequestFailure); ok { + // A service error occurred + errType = errType + fmt.Sprintf(" %d %s", reqErr.StatusCode(), aerr.Code()) + } else { + errType = errType + fmt.Sprintf(" 000 %s", aerr.Code()) } - // Other 404 errors like NoSuchVersion and - // NoSuchBucket are different problems which should - // get called out downstream, so we don't convert them - // to os.ErrNotExist. } - return err -} - -// EmptyTrash looks for trashed blocks that exceeded trashLifetime -// and deletes them from the volume. -// TBD -func (v *S3Volume) EmptyTrash() { + s.statsTicker.TickErr(err, errType) }